Simple Text Classification with BERT-avgpool
This is actually the code for two projects, one on data augmentation, which I wrote a paper for, and one on gradient mining, which didn't work out because it was actually a long winded way of using a MLP.
code/{1,2,3} is all for data augmentation. Most of the other code is for gradient mining.
You can get the batch-level gradient directly from PyTorch.
For getting per-example gradients, the only library that I know of that works is autograd-hacks.